Estimation of sunlight direction using 3D object models.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Sunlight direction is crucial for image processing tasks like augmented reality and object recognition.
• Existing methods often rely on weak cues (sky, texture, shadows) and may require known camera parameters, limiting their use.
• A robust method for estimating sunlight direction from single images is needed.

Purpose of the Study:

• To develop a novel method for estimating sunlight direction from a single outdoor image.
• To overcome limitations of existing methods by using object recognition and shadow inference.
• To improve the accuracy and applicability of sunlight direction estimation.

Main Methods:

• Object recognition using exemplar-SVMs trained on 2D samples from 3D models.
• Inferring cast shadows from recognized objects using 3D object models.
• Estimating sunlight direction by analyzing inferred shadows under various illumination directions.

Main Results:

• Accurate estimation of the azimuth angle within a quadrant (error < 45°).
• Precise computation of the zenith angle with a mean angular error of 23°.
• Demonstrated effectiveness on both synthetic and real-world images.

Conclusions:

• The proposed method effectively estimates sunlight direction from single images.
• Object modeling and shadow inference provide reliable cues for solar angle estimation.
• This approach enhances applications requiring accurate sunlight direction information.